Blog

What to do while the AI writes code

Agents make you faster. Filling that speed with more loops burns you out. Use the wait to read, talk to someone, and write what you shipped.

An agent can put a pull request up before you finish reading the ticket. A page that used to take a day shows up before lunch. The obvious next move is to open another agent and do it again.

That is how a fast week turns into a pile of work nobody remembers.

Five times faster, ten times tired

The first stretch feels like a gift. Days of work collapse into an afternoon. You take the infrastructure bug, the settings page, and the admin tool at the same time. Tickets write themselves. Threads get replies. The board looks full, and for a week you believe the job got bigger because you got faster.

Then the context falls out. Pull requests sit for weeks because the reason they exist left with the chat. Reviews get thinner because you are already in the next loop. You are not holding five projects. You are skimming five of them, and the one that ships is the one you happened to still have open.

You did not get a bigger product. You got tired faster, and the diff got noisier. More parallel work is not what the extra time is for.

Read past the snippet

Most of us learned to skim. Find the code block, take the conclusion, leave the essay. That made sense when typing the implementation was the slow part of the day. You could not afford someone else’s long explanation when the build was still yours to write.

It is not the slow part anymore. While the agent runs, read the page you would have searched. Not the snippet. The part where the author says what broke, what they tried first, and what they would not do again. Docs reward the same patience. Learn how the API actually behaves before you ask for a patch.

A React View Transition is a small case. The snippet gets you a demo. The page around it tells you when the transition lies, which state it fails to capture, and what you should not ask a model to invent. That is the difference between a feature you can explain and a feature you can only regenerate.

That reading changes the prompt. You stop asking the agent to discover the feature and start telling it the path you would have taken. The loop gets shorter because you already know what done means, and you can reject the first answer that merely compiles.

Ask someone to look with you

The other use of the wait is a person. Message a teammate and look at the change together. Not a status update in the thread the agent already replied to. The edge cases, the naming, the part of the product you have not sat with in months.

An agent can summarize a folder in a minute. It cannot tell you which compromise the last person already tried and rejected, or why a name that looks clean will confuse the people who support the product. That lives in a conversation. If you only ever hand the next agent a transcript, you will walk the dead ends again, faster, and call it progress.

Pairing still does the job it always did. It is how a decision leaves your head and becomes something another human can defend when you are not in the room.

Write the note

The best use of those minutes is writing.

Fifteen minutes. What changed, who it is for, and what you are still unsure about. Before agents, that felt like time you did not have. The calendar was full of the typing. Now the typing is cheap, and the note is the part of the job a model cannot do honestly for you. It does not know which sentence you almost cut, or which customer complaint the change is actually answering.

People read those notes. They comment. The next feature often starts in a sentence you almost skipped: a rule you had to explain, a page someone asked for, a date you put on a changelog so the release could be found again. A private chat log does not do that. A public URL does.

ProductPost 360 starts from that note — a changelog entry, a blog post, a page that already exists. We draft what each place expects. You still paste it. You stay in the thread. The writing has to be yours, because that is what another person will trust enough to link, and what Google can still find a month later.

Fewer ships, clearer ones

The backlog does not shrink because you got faster. Someone will remember the thing you shipped in a day and ask for the same pace next week. The feed is full of side projects launched the same afternoon. The pull is to chase that feeling and open one more loop.

Careful work compounds. Think about the decision before you accept the diff. Write it down where a customer can read it. Talk to someone who uses the product, not only to the agent that just edited it. You will not ship ten times as many things. The things you do ship will still have a URL next month, and a reason you can say out loud.

More notes

All notes

Legaltech where firms already talk

Law-firm, contract, and compliance tools now map to ALCA, Docusign, and Acrobat Sign — plus the legal subreddits, not a generic SaaS board.

The pages answer engines cite instead of you

A mention is not the same as a citation. See which domains the answer used, and which of those pages never name your product.

What happened after the click

Analytics shows who arrived, which channel sent them, and which referrer they came from. Search Console never sees that half of the visit.

Searches that just showed up

Search Console on the performance page shows queries that appeared in this 28-day window, the pages behind them, and the days you posted drafts.

Four sources, one brief

Search Console, Bing, Analytics, and backlinks each leave a blind spot. One brief turns those snapshots into recommendations and opportunities.

How often answer engines mention you

Gemini, ChatGPT, and Perplexity do not share one ranking. Track how often each names your product on the questions buyers actually ask.